KAIST XAI Tutorial Series (KAIST XAI 튜토리얼 시리즈)
Introduction
- Target audience:
- The newcomers to the SAIL lab, PhD and MSc students who are interested
- Members from collaboration institutions who are interested
- When: Jan. 26(Thu) ~ Feb. 16(Tue), every Thursday&Tuesday 16:00-18:00
- Where: Hybrid, KAIST XAI Research Center and Zoom
- Each Session: 60min (including 5~10min for Q&A)
- Language: Korean/English (Slides in English)
Program
DAY 1. Jan 26 (Thu)
1. Recent Trends in Explainable Artificial Intelligence
Jaesik Choi (Professor, KAIST AI)
As complex artificial intelligence (AI) systems such as deep neural networks is
used for many mission critical task such as military, finance, human resources
and autonomous driving, it is important to secure the safe use of such complex
AI systems. In this talk, we will overview the principles and recent advances in
explainable artificial intelligence.
2. Symbolic AI methods
Seongwoo Lim (SAILab, KAIST AI)
Symbolic representation can be interpretable to both human and computer. In this
session, we will introduce symbolic AI models like decsion tree and
(first-order) logic based models. Also, we will introduce neuro-symbolic models,
which combine neural network and symbolic model.
DAY 2. Jan 31 (Tue)
3. Global model-agnostic methods
Wonjoon Chang (SAILab, KAIST AI)
Global interpretation methods describe the average behavior of data or a model.
They are particularly useful when a user wants to understand the general
mechanisms in the data or debug the model.
4. Local model-agnostic methods
Jiyeon Han (SAILab, KAIST AI)
In this session, we discuss about model-agnostic methods to explain individual
predictions. In specific, we mainly deal with two approaches, LIME and Shapley
values.
DAY 3. Feb 02 (Thu)
5. CAM-based methods
Jungmin Kim (SAILab, KAIST AI)
Class Activation Maps(CAM) is motivated by CNN's ability to localize the
discriminative regions for classification. XAI methods based on Activation
(Grad-CAM, Grad-CAM++, etc.) have been developed to improve the limiations of
previous activation methods. In this session, we explore the motivation of CAM
based methods and how the other methods are developed.
6. Input attribution methods
Seongun Kim (SAILab, KAIST AI)
Input attribution methods explain the model through quantifying importance score
of input features for the model. In this session, we mainly deal with three
input attribution methods, LRP, RAP, and IG.
DAY 4. Feb 07 (Tue)
7. Perturbation-based explanations
Sehyun Lee (SAILab, KAIST AI)
The talk will focus on two important perturbation methods of Explainable AI
(XAI): RISE and SISE. RISE (Random Interpolation of Sensitized Examples)
generates saliency maps to highlight the areas of an input that have the most
impact on the model's prediction. While Similar with RISE, SISE (Semantic Input
Sampling for Explanation) uses a guided sampling approach to generate
explanations. Both RISE and SISE are discussed in detail, with examples to
illustrate how they can be used to interpret the predictions made by AI models.
8. Concept-based explanations
Dahee Kwon (SAILab, KAIST AI)
In this session, we will discuss concept-based explanations which provide
human-interpretable explanations. We can measure the importance of each concept,
(e.g, how much influence the stripe pattern has on zebra prediction). Also, we
can visualize the important learned concepts from the model.
DAY 5. Feb 09 (Thu)
9. Attention as Explanation
Cheongwoong Kang (SAILab, KAIST AI)
This session includes two sub-topics: (1) how attention is used as local/global
explanations and (2) quantifying attention flow in transformers, where
information from different tokens gets increasingly mixed.
10. Evaluation of XAI
Junho Choi (SAILab, KAIST AI)
The definition of "good explanation" can vary by audience and problem: it is
subjective and diverse. Nevertheless, there are some qualities that one could
expect from a "good explanation". This presentation will go over Quantus, an
open-source package that lists these qualities and the methods for their
quantitative evaluation.
DAY 6. Feb 14 (Tue)
11. Application of XAI Toolkit for PyTorch
Soyeon Kim (SAILab, KAIST AI)
In this tutorial we use Captum library to explain pre-trained models with
various input attribution methods which we studied previous session: Saliency,
Grad-CAM, LRP, IG, perturbation-based, Shapley values.
12. Explanation tools of Google What-If & LIT
Seongyeop Jeong (SAILab, KAIST AI)
We will talk about google explanation tools which serve automatic explanation
algorithm IG, Shapley value, visualization of features, and helpful interface.
The complex model could be easily analyzed by these tools with the intuitive
interface representing which information is important & model's result is
trustable.
DAY 7. Feb 16(Thu)
13. Tutorial on Chain of Thoughts (English)
Minjoon Seo (Professor, KAIST AI)
As complex artificial intelligence (AI) systems such as deep neural networks is
used for many mission critical task such as military, finance, human resources
and autonomous driving, it is important to secure the safe use of such complex
AI systems. In this talk, we will overview the principles and recent advances in
explainable artificial intelligence.
Reading List:
14. Human factors engineering for Explainable AI (English)
Woojin Park (Professor, Seoul National University)
Symbolic representation can be interpretable to both human and computer. In this
session, we will introduce symbolic AI models like decsion tree and
(first-order) logic based models. Also, we will introduce neuro-symbolic models,
which combine neural network and symbolic model.